UAV-based fusion SLAM measurement method for underground goaf areas

Abstract

This invention provides a UAV-based fusion SLAM measurement method for underground mining areas, comprising: controlling a UAV to enter the underground mining area, simultaneously acquiring laser point cloud data, visual image data, and inertial attitude data through a lidar, visual camera, and inertial measurement unit, and forming a fusion measurement data sequence after time synchronization, coordinate transformation, and validity screening; performing robust pose estimation based on the fusion measurement data sequence, generating robust pose trajectories and constructing local map units, and generating suspected closed-loop records through a machine learning similarity discrimination model; performing lateral scanning, pitch scanning, or reverse look-back measurements at the suspected closed-loop locations to form verification measurement segments, and matching them with historical measurement segments to generate verification matching results; constructing a closed-loop counter-evidence chain based on the verification matching results to determine whether the suspected closed-loop location is a valid closed loop or a pseudo-similar closed loop; and generating valid closed-loop constraints or performing local pose correction based on the closed-loop determination results.

Description

Technical Field

[0001] This invention relates to the field of underground space measurement and navigation positioning technology, and in particular to a UAV-based fusion SLAM measurement method for underground mining areas. Background Technology

[0002] Underground goaf areas typically exhibit conditions such as lack of GNSS signals, irregular spatial structures, and localized collapses. Existing technologies often employ unmanned aerial vehicles (UAVs) equipped with lidar, visual cameras, and inertial measurement units (IMUs) to enter the goaf area. By integrating SLAM (Simultaneous Localization and Mapping), the UAVs achieve localization, point cloud registration, and 3D map construction, acquiring goaf boundaries, cross-sectional contours, cavity volumes, and coordinates of hazardous areas.

[0003] Existing fusion SLAM measurement methods rely on laser point cloud registration, visual feature matching, and inertial attitude prediction for pose estimation. They generate closed-loop constraints when the current location and historical locations have high similarity to reduce trajectory drift. However, underground goaf areas often contain similar mining cross-sections, similar sidewall curvatures, and overlapping roof and floor structures. The point cloud distribution and visual features acquired from different spatial locations may be highly similar. The system is prone to misclassifying two actually different locations as the same location and using this misclassification as an effective closed-loop constraint in global map optimization.

[0004] Even after error-induced loop closures are incorporated into optimization, the generated trajectories and point clouds may still appear continuous, smooth, and have low residuals. This results in a seemingly self-consistent 3D map at the data level, but in reality, spatial folding, positional mismatches, or local error closures have occurred. Such problems affect the accuracy of goaf boundaries, cross-sectional contours, cavity volumes, and hazardous area coordinates. Therefore, this invention proposes a UAV-based fusion SLAM measurement method for underground goaf areas.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may contain prior art information that is not common knowledge to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a UAV-based fusion SLAM measurement method for underground mining areas, thereby solving the technical problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A UAV-based fusion SLAM measurement method for underground goaf areas includes the following steps: S1. Control the UAV to enter the underground mining area and simultaneously collect laser point cloud data, visual image data and inertial attitude data through lidar, vision camera and inertial measurement unit. After time synchronization, coordinate transformation and validity screening, a fused measurement data sequence is formed. S2. Based on the fused measurement data sequence, robust pose estimation is performed to generate robust pose trajectory, and local map units are constructed. The cross-sectional contour, relative height of top and bottom plates, side curvature and local visual features are extracted, and suspected closed-loop records are generated by machine learning similarity discrimination model. S3. Determine the safe flight boundary of the UAV based on the suspected closed-loop record, establish a redundancy verification mechanism based on real-time perception data, generate verification flight action commands, control the UAV to perform lateral scanning, pitch scanning or reverse review, form verification measurement segments, and match them with historical measurement segments to generate verification matching results. S4. Generate closed-loop counter-evidence factors based on the verification matching results, construct closed-loop counter-evidence chains, and output the closed-loop determination results of valid closed loops or pseudo-similar closed loops based on the closed-loop counter-evidence chains. S5. Generate effective closed-loop constraints or not generate corresponding closed-loop constraints based on the closed-loop determination results. Use the effective closed-loop constraints, local map units and verification measurement segments to correct the robust pose trajectory, obtain the high-precision positioning trajectory and the corrected three-dimensional map of the goaf, and generate the fusion SLAM measurement results of the goaf.

[0008] S1 specifically includes: reading the preset measurement path, entrance reference coordinates, UAV initial attitude, and installation parameters of each sensor; using the sampling time of the inertial measurement unit as the main time axis, synchronizing the laser point cloud data, visual image data, and inertial attitude data to form initial synchronized acquisition data; establishing a measurement coordinate system for the underground goaf area based on the entrance reference coordinates and the UAV initial attitude; converting the laser point cloud data to this coordinate system; and removing invalid sensor frames, frames with missing timestamps, and abnormal data to obtain valid synchronized measurement data; encapsulating the valid synchronized measurement data according to a unified time identifier; recording the UAV sampling position, point cloud density, image clarity, inertial attitude change, data status markers, and sensor reliability to form a fused measurement data sequence.

[0009] S2 specifically includes: taking the fused measurement data sequence as input, reading laser point cloud data, visual image data, inertial attitude data, UAV sampling position and sensor reliability, performing inertial prediction, point cloud registration correction and visual feature tracking correction, generating robust pose trajectory, and constructing local map units; The cross-sectional contour, relative height of top and bottom plates, cross-sectional width, sidewall curvature, tunnel width-to-height ratio, local visual features, and laser point cloud distribution features are extracted from local map units, input into a machine learning similarity discrimination model, and combined with weight formulas to obtain structural similarity and feature credibility. Non-adjacent local map units are used as candidate comparison objects. When the structural similarity, feature credibility, and path interval all meet the threshold conditions, a suspected closed loop record is generated.

[0010] S3 specifically includes: after generating a suspected closed-loop record, reading the current measurement segment, historical measurement segment, current robust pose, historical robust pose and suspected closed-loop triggering cause, combining the cross-section width, relative height of the top and bottom plates, side curvature, local visual feature matching distribution and laser point cloud set to determine the safe flight boundary of the UAV, and combining the real-time ranging data of the UAV to construct a redundant verification mechanism for collision avoidance protection, and generating verification flight action commands; The UAV is controlled to perform short-range measurements according to the verification flight action commands. During the execution, the obstacle avoidance distance is monitored in real time through a redundancy verification mechanism. Simultaneously, verification laser point cloud data, verification visual image data, and verification inertial attitude data are collected and packaged into verification measurement segments according to a unified time identifier. The verification measurement segments are then compared with historical measurement segments according to cross-sectional position, sampling direction, and action type to establish a correspondence of similar features and generate verification matching results.

[0011] S4 specifically includes: taking the verification matching results as input, reading the sequence of cross-section changes, the difference in the relative height of the top and bottom plates, the change in sidewall curvature, the cumulative amount of inertial heading, the intersection relationship of laser rays, the visual feature review of matching data and the matching credibility, and generating closed-loop counter-evidence factors; performing consistency checks on each closed-loop counter-evidence factor according to a preset tolerance threshold, determining the counter-evidence state, and forming a closed-loop counter-evidence chain according to the sequence of cross-section changes, the relative height of the top and bottom plates, the sidewall curvature, the cumulative amount of inertial heading, and the intersection relationship of laser rays; calculating the comprehensive counter-evidence score based on the closed-loop counter-evidence chain, and determining it as a pseudo-similar closed loop when a strong counter-evidence factor is established or the comprehensive counter-evidence score reaches the threshold, otherwise determining it as a valid closed loop, and generating a closed-loop judgment result.

[0012] S5 specifically includes: processing suspected closed-loop records based on the closed-loop determination results; generating effective closed-loop constraints and incorporating them into global map optimization when a valid closed loop occurs; not generating closed-loop constraints when a pseudo-similar closed loop occurs; and retaining the verification measurement segments and local pose correction range markers. Using the robust pose trajectory as the object to be corrected, global map optimization and local pose correction are performed using effective closed-loop constraints, local map units, verification measurement segments, and verification matching results to obtain a high-precision positioning trajectory and a corrected 3D map of the goaf. Based on the high-precision positioning trajectory and the corrected 3D map of the goaf, the goaf boundary, cross-sectional contour, cavity volume, and hazardous area coordinates are extracted, and associated with the closed-loop determination results, effective closed-loop constraints, suspected closed-loop records judged as pseudo-similar closed loops, verification measurement segment indexes, and local pose correction records to form a fused SLAM measurement result for the goaf.

[0013] The beneficial effects of this invention are as follows: This invention achieves time synchronization, coordinate transformation, and validity screening of laser point cloud data, visual image data, and inertial attitude data to form a fused measurement data sequence. This provides underground goaf measurement data with a unified time and spatial reference, improving the reliability of subsequent robust pose estimation. Suspected loop closures are identified using sensor reliability, robust pose trajectories, and local map units. A machine learning similarity discrimination model is then used to filter suspected loop closure locations, enhancing the targeting of loop closure candidate identification in repetitive roadways, similar cross-sections, and weak texture environments.

[0014] This invention generates verification measurement segments by triggering lateral scans, pitch scans, or reverse look-back measurements at suspected closed-loop locations. These segments are then matched with historical measurement segments for similar features. This proactively verifies the authenticity of closed loops before they are incorporated into global map optimization, reducing the risk of false closed loops. By constructing a closed-loop counter-verification chain using the sequence of cross-sectional changes, the relative height of the top and bottom plates, sidewall curvature, cumulative inertial heading, and laser ray intersection relationships, it can identify pseudo-similar closed loops with similar structures but different spatial locations, preventing seemingly self-consistent but actually erroneous spatial folds in the map.

[0015] This invention generates valid closed-loop constraints when effective closed-loop occurs, but does not generate corresponding closed-loop constraints when pseudo-similar closed-loop occurs, and instead uses verified measurement segments for local pose correction. This suppresses the impact of erroneous closed-loops on global map optimization, improving the accuracy of high-precision positioning trajectories and corrected 3D maps of goaf areas. Based on the high-precision positioning trajectory and corrected 3D maps of goaf areas, the invention outputs goaf boundaries, cross-sectional contours, cavity volumes, and hazardous area coordinates, enabling the measurement results to more accurately serve goaf management, backfilling design, and hazardous area delineation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the UAV-based fusion SLAM measurement method for underground mining areas according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example: Figure 1 As shown, this embodiment provides a UAV-based fusion SLAM measurement method for underground goaf areas, including the following steps: S1. Control the UAV to enter the underground mining area and simultaneously collect laser point cloud data, visual image data and inertial attitude data through lidar, vision camera and inertial measurement unit. After time synchronization, coordinate transformation and validity screening, a fused measurement data sequence is formed. S2. Based on the fused measurement data sequence, robust pose estimation is performed to generate robust pose trajectory, and local map units are constructed. The cross-sectional contour, relative height of top and bottom plates, side curvature and local visual features are extracted, and suspected closed-loop records are generated by machine learning similarity discrimination model. S3. Determine the safe flight boundary of the UAV based on the suspected closed-loop record, establish a redundancy verification mechanism based on real-time perception data, generate verification flight action commands, control the UAV to perform lateral scanning, pitch scanning or reverse review, form verification measurement segments, and match them with historical measurement segments to generate verification matching results. S4. Generate closed-loop counter-evidence factors based on the verification matching results, construct closed-loop counter-evidence chains, and output the closed-loop determination results of valid closed loops or pseudo-similar closed loops based on the closed-loop counter-evidence chains. S5. Generate effective closed-loop constraints or not generate corresponding closed-loop constraints based on the closed-loop determination results. Use the effective closed-loop constraints, local map units and verification measurement segments to correct the robust pose trajectory, obtain the high-precision positioning trajectory and the corrected three-dimensional map of the goaf, and generate the fusion SLAM measurement results of the goaf.

[0019] S1 specifically includes the following sub-steps: S110. After controlling the UAV to enter the entrance of the underground goaf area, read the preset measurement path, entrance reference coordinates, UAV initial attitude, lidar installation parameters, vision camera installation parameters, and inertial measurement unit installation parameters. The entrance reference coordinates are derived from existing measurement control points in the mine, entrance control points determined by the total station, or entrance calibration points set up before operation. The initial attitude of the UAV is derived from the attitude calculation results output by the inertial measurement unit in a stationary state before takeoff. The lidar installation parameters and vision camera installation parameters are derived from factory calibration or pre-operation calibration.

[0020] Using the sampling time of the inertial measurement unit as the main time axis, the laser point cloud data output by the lidar, the visual image data output by the vision camera, and the inertial attitude data output by the inertial measurement unit are read respectively, and time synchronization is performed according to the timestamp carried by each data frame.

[0021] For any single frame of laser point cloud or visual image, calculate the time difference between its sampling time and the corresponding inertial sampling time:

[0022] in, This represents the time difference between the sensor data and the inertial sampling time. This indicates the sampling time of a laser point cloud frame or a visual image frame. This indicates the sampling time of the inertial measurement unit.

[0023] when When the value is not greater than the preset synchronization threshold, the corresponding laser point cloud data, visual image data, and inertial attitude data are grouped into the same sampling time; when... When the value exceeds the preset synchronization threshold, the corresponding data frame is marked as a time-asynchronous frame and is not included in the data set at that sampling time. This forms the initial synchronized acquisition data; the initial synchronized acquisition data refers to the data set composed of laser point cloud data, visual image data, and inertial attitude data at the same sampling time, which is used to subsequently establish the measurement coordinate system of the underground goaf area and generate valid synchronized measurement data.

[0024] In one specific embodiment, the preset synchronization threshold is set to 10 milliseconds. In the sensor confidence calculation of S130, to adapt to the environment of underground mining areas with high dust levels but still acceptable texture, the weighting coefficient can be set as: laser point cloud confidence weight. =0.4, visual image credibility weight =0.4, Inertial Attitude Confidence Weight =0.2.

[0025] S120. Establish a measurement coordinate system for the underground goaf area based on the entrance reference coordinates and the initial attitude of the UAV, and use the UAV body coordinate system as an intermediate coordinate system to transform the laser point cloud data from the lidar coordinate system to the measurement coordinate system for the underground goaf area.

[0026] Specifically, the laser point cloud data is first converted to the UAV body coordinate system based on the lidar installation parameters, and then converted to the underground goaf measurement coordinate system based on the entrance reference coordinates and the UAV's initial attitude. The coordinate transformation relationship is as follows:

[0027] in, This represents the coordinates of a point in the point cloud in the lidar coordinate system. This represents the coordinates of a point cloud in the coordinate system used for measuring underground goaf areas. This represents the rotation matrix from the lidar coordinate system to the UAV body coordinate system. This represents the translation vector from the lidar coordinate system to the UAV body coordinate system. This represents the rotation matrix from the UAV's body coordinate system to the measurement coordinate system of the underground goaf area. This represents the translation vector from the UAV body coordinate system to the measurement coordinate system of the underground mining area.

[0028] Visual image data and inertial attitude data are mapped according to a unified time identifier, and the validity of the initial synchronously acquired data is screened. The validity screening includes removing invalid sensor frames, frames with missing timestamps, and abnormal data. Invalid sensor frames include laser point cloud frames with fewer than a preset number of valid points and visual image frames with abnormal exposure or lower than a preset clarity threshold. Abnormal data includes data frames whose angular velocity or acceleration calculated from inertial attitude data exceeds the maximum angular velocity or maximum acceleration constraint of the UAV.

[0029] For example, if the maximum permissible angular velocity of a UAV is 120 deg / s, but the angular velocity corresponding to the attitude angle change calculated at two adjacent inertial sampling times is 300 deg / s, then this inertial attitude data is judged as abnormal data and discarded. After coordinate transformation, time correspondence, and validity screening, valid synchronized measurement data is obtained. Valid synchronized measurement data refers to measurement data that still meets the time synchronization and coordinate transformation requirements after removing invalid frames, frames with missing timestamps, and abnormal data, and is used to subsequently form a fused measurement data sequence.

[0030] S130. Encapsulate the valid synchronous measurement data according to a unified time identifier to form a fused measurement data sequence arranged in chronological order.

[0031] Each set of fused measurement data includes laser point cloud data in the underground goaf measurement coordinate system, visual image data corresponding to the sampling time, inertial attitude data, UAV sampling position, laser point cloud density, visual image sharpness, inertial attitude change, data status markers, and sensor reliability. The fused measurement data sequence refers to multiple sets of fused measurement data arranged in a unified time identifier order, which are used for subsequent robust pose estimation, local map unit construction, suspected loop closure identification, and high-precision positioning correction.

[0032] Sensor reliability is calculated based on laser point cloud reliability, visual image reliability, and inertial attitude reliability.

[0033] Where C represents the sensor reliability of a set of fused measurement data. Indicates the credibility of laser point clouds. Indicating the credibility of visual images, Indicates the reliability of inertial attitude. , , These represent the weighting coefficients for the credibility of the laser point cloud, the credibility of the visual image, and the credibility of the inertial attitude, respectively, and the sum of the three weighting coefficients is 1.

[0034] The reliability of laser point clouds is determined based on the number of effective points per unit space and the continuity of point cloud distribution. The reliability of visual images is determined based on image sharpness and the number of trackable feature points. The reliability of inertial attitude is determined based on the continuity of attitude changes between adjacent sampling times. Robustness in robust pose estimation refers to maintaining the continuity of UAV pose estimation through sensor reliability adjustment and outlier data removal when faced with dust occlusion, weak texture, sparse point clouds, or degraded data quality from a single sensor.

[0035] Therefore, the fused measurement data sequence not only records the synchronous measurement results of multiple source sensors, but also records the data quality and reliability status at each sampling moment, enabling S210 to generate a robust pose trajectory based on the fused measurement data sequence, and enabling S220 and S230 to use the laser point cloud density, visual image clarity and sensor reliability to filter local map units and suspected closed loop locations.

[0036] S2 specifically includes the following sub-steps: S210: Using the fused measurement data sequence formed in S130 as input, the laser point cloud data, visual image data, inertial attitude data, UAV sampling position and sensor reliability are read frame by frame in the underground goaf measurement coordinate system according to the unified time identifier, and the UAV continuous pose estimation is performed.

[0037] Specifically, the attitude change and acceleration change in the inertial attitude data are used as the basis for short-term motion prediction to obtain the predicted pose at the current sampling time. Laser point cloud data from adjacent sampling times are registered to obtain the laser point cloud registration error. Trackable feature points are matched with visual image data from adjacent sampling times to obtain the visual feature tracking error. Then, combined with the inertial attitude prediction error, the predicted pose is weighted and corrected to form the robust pose at the current sampling time. Robust pose refers to the UAV pose that can still be output after adjusting the error weights based on sensor reliability even when the quality of any type of data—laser point cloud data, visual image data, or inertial attitude data—deteriorates.

[0038] The pose estimation error is calculated according to the following formula:

[0039] Where E represents the pose estimation error at the current sampling time. This indicates the laser point cloud registration error. Indicates visual feature tracking error. This represents the inertial attitude prediction error. The weights representing the laser point cloud registration error are: The weights representing the visual feature tracking error, This represents the weight of the inertial attitude prediction error.

[0040] The aforementioned weights are determined by the sensor reliability in S130; when dust causes a decrease in laser point cloud density, the weights are reduced. When weak textures lead to a reduction in traceable feature points, the reduction is achieved. When the inertial attitude change is continuous, improve The constraint effect in short-term forecasting.

[0041] Robust pose sequences are generated by connecting the robust poses output at consecutive sampling times in chronological order. A robust pose sequence refers to a continuous pose sequence of a UAV obtained based on a fused measurement data sequence, after inertial attitude prediction, laser point cloud registration and correction, visual feature tracking and correction, and sensor confidence weighting.

[0042] Subsequently, using the robust pose trajectory as an index, local map units are divided according to preset flight distance windows, preset time windows, or cross-sectional change events. Each local map unit includes at least a robust pose trajectory segment within the corresponding window, a set of laser point clouds, a set of visual image features, the cross-sectional location, and sensor confidence statistics. Local map units are used for subsequent extraction of cross-sectional structural features and serve as the basic comparison objects for suspected loop closure identification.

[0043] S220. For each local map unit, determine the UAV's travel direction based on the robust pose trajectory segment, and extract the laser point cloud set in a direction approximately perpendicular to the UAV's travel direction to form a goaf cross-section point cloud; extract the cross-section contour, relative height between the roof and floor, cross-section width, sidewall curvature, and roadway width-to-height ratio based on the effective boundary points of the roof, floor, left sidewall, and right sidewall in the cross-section point cloud.

[0044] The relative height between the top and bottom plates is determined by the height difference between the effective boundary points of the top plate and the bottom plate within the same cross-section. The cross-sectional width is determined by the lateral distance between the effective boundary points of the left and right sides within the same cross-section. The side curvature is determined by the continuous variation trend of the side boundary points along the cross-sectional direction.

[0045] Specifically, when extracting the cross-sectional profile and sidewall curvature, voxel filtering and the Random Sample Consensus (RANSAC) algorithm are first used to remove suspended noise points caused by dust. Then, the effective boundary points are projected onto the two-dimensional cross-sectional plane, and the least squares method is used to perform polynomial curve fitting. Finally, the reciprocal of the radius of curvature is calculated as the sidewall curvature value.

[0046] The width-to-height ratio of the tunnel is calculated using the following formula:

[0047] in, Indicates the width-to-height ratio of the alleyway. Indicates the cross-sectional width. This indicates the relative height between the top and bottom plates.

[0048] Simultaneously, stably trackable corner points, edge points, and texture blocks are extracted from the visual image feature set to form local visual features; the number of effective points per unit space, point cloud continuity, and occlusion gap locations are statistically analyzed from the laser point cloud set to form laser point cloud distribution features. The differences in cross-sectional contours, relative heights of the top and bottom plates, sidewall curvatures, tunnel width-to-height ratios, number of matched local visual features, laser point cloud distribution similarity, and sensor reliability of the two local map units to be compared are input into a trained machine learning similarity discrimination model, which outputs cross-sectional contour similarity. Relative height similarity of top and bottom plates Side curvature similarity Local visual feature similarity and feature credibility .

[0049] The machine learning similarity discrimination model specifically adopts the Siamese Network architecture. This model contains two feature extraction sub-networks with shared weights, each consisting of three fully connected layers (multilayer perceptrons) with ReLU activation function.

[0050] The model takes a one-dimensional vector as input, containing features such as cross-sectional contour, top and bottom plate height, and sidewall curvature, and outputs a high-dimensional feature representation. During training, a contrastive loss function is used, ensuring that the distance between effective closed-loop sample pairs (positive samples) at the same spatial location is less than a preset margin in the high-dimensional space, while the distance between pseudo-similar sample pairs (negative samples) at different spatial locations is greater than a preset margin.

[0051] After the model is trained and converged using the backpropagation algorithm, it is used for similarity discrimination. In a specific embodiment, the weight coefficient in the structural similarity calculation formula can be set as: cross-sectional contour similarity weight. =0.35, relative high similarity weight =0.25, curvature similarity weight =0.20, visual feature similarity weight =0.20.

[0052] Structural similarity is calculated using the following formula:

[0053] Where S represents the structural similarity between two local map units. Indicates the similarity of cross-sectional profiles. Indicates the similarity of the relative heights of the top and bottom plates. Indicates the similarity of sidewall curvature. Indicates the similarity of local visual features. , , , These represent the weight coefficients for the corresponding similarity scores, and the sum of the four weight coefficients is 1. Structural similarity and feature confidence are used for subsequent screening of suspected loop closure locations.

[0054] S230. Using any two non-adjacent local map units as candidate comparison objects, based on the structural similarity and feature reliability output in S220, and combined with the path interval of the two local map units on the robust pose trajectory, determine whether a suspected loop closure record is generated. To avoid misclassifying naturally similar adjacent sections as loop closures during continuous flight, the suspected loop closure judgment is only allowed after the two local map units meet the path interval condition.

[0055] The criteria for judging a suspected closed loop are:

[0056] Where S represents the structural similarity between two local map units. This indicates a preset similarity threshold. Indicates the credibility of features. This indicates a preset confidence threshold. This represents the path interval between two local map cells on the robust pose trajectory. This indicates the preset path interval threshold.

[0057] When the structural similarity is not lower than the preset similarity threshold, the feature credibility is not lower than the preset credibility threshold, and the path interval is not lower than the preset path interval threshold, the two local map units are identified as suspected closed loop locations and suspected closed loop records are generated; otherwise, no suspected closed loop records are generated.

[0058] A suspected loop closure record refers to a set of candidate loop closure information formed after two local map units meet the conditions of structural similarity, feature confidence, and path interval. It includes at least the current measurement segment, historical measurement segments, current robust pose, historical robust pose, structural similarity, feature confidence, path interval, and the suspected loop closure trigger reason. Here, the current measurement segment originates from the local map unit currently traversed by the UAV, and the historical measurement segments originate from historical local map units established in the robust pose trajectory.

[0059] For example, if a drone passes through a local area with a cross-sectional width of 4m and a relative height between the top and bottom plates of 3m from 35s to 38s, and then passes through a local area with a cross-sectional width of 4.1m and a relative height between the top and bottom plates of 3.05m from 118s to 121s, a suspected loop closure record is generated if the structural similarity, feature reliability, and path interval of the two areas all meet the above conditions. This suspected loop closure record is only used for subsequent S310-triggered lateral scans, pitch scans, or reverse lookback measurements, and is not directly added to the global map optimization as an effective loop closure constraint.

[0060] In a practical goaf measurement embodiment, the preset parameters for determining suspected closed loops are set as follows: the preset similarity threshold S0 is set to 0.85, and the preset confidence threshold... The path interval threshold is set to 0.75 to avoid mismatches between adjacent sections. It is set to 30 meters.

[0061] S3 specifically includes the following sub-steps: S310. After generating a suspected closed-loop record in S230, read the current measurement segment, historical measurement segment, current robust pose, historical robust pose, and suspected closed-loop triggering reason from the suspected closed-loop record. Combined with the cross-sectional width, relative height between the top and bottom plates, sidewall curvature, local visual feature matching distribution, and laser point cloud set in the current measurement segment, determine the safe flight boundary of the UAV near the suspected closed-loop location. The safe flight boundary of the UAV refers to the minimum usable distance between the current position of the UAV and the left and right sidewalls, top plate, and bottom plate of the goaf, used to limit the execution range of verification measurement actions.

[0062] The safe flight boundary for drones is calculated using the following formula:

[0063] in, Indicates the safe flight boundary of the drone. This indicates the distance from the drone's current position to the left side. This indicates the distance from the drone's current position to the right side of the wall. This indicates the distance from the drone's current position to the top panel. This indicates the distance from the drone's current position to the ground. The distances mentioned above are determined by the current robust pose and the laser point cloud data in the current measurement segment.

[0064] After determining the safe flight boundary of the UAV based on static point clouds, the system establishes a collision avoidance redundancy verification mechanism based on real-time sensor data to provide both theoretical and physical collision avoidance protection. The specific logic of this redundancy verification mechanism is as follows: During the system's generation and execution of verification actions, real-time high-frequency detection data from lidar or vision cameras are used as dynamic obstacle avoidance input; if the calculated safe flight boundary... If the distance to the nearest obstacle is greater than the preset safe distance and the real-time distance to the nearest obstacle during the verification flight is never lower than the preset emergency braking threshold, the current environment is deemed safe and the verification action command continues to be executed. If the real-time distance to the nearest obstacle is triggered by local unknown protrusions or sudden dust in the goaf, causing the preset emergency braking threshold to be triggered, the redundancy verification mechanism will immediately forcefully interrupt the current verification flight action command and control the UAV to hover in place.

[0065] This mechanism overcomes the limitations of static boundary prediction and ensures the absolute safety of UAVs when performing verification scanning in complex goaf environments.

[0066] when When the distance is not less than the preset safe distance of the UAV and the suspected closed-loop triggering cause is mainly due to similar curvature or cross-sectional width of the left and right sidewalls, a transverse scan command is generated. The transverse scan command means controlling the UAV to move a short distance laterally along the current cross-section while keeping the heading unchanged, in order to obtain data on the changes in curvature and cross-sectional width of the left and right sidewalls.

[0067] when When the distance is not less than the preset safe distance of the UAV, and the suspected closed-loop triggering cause is mainly due to the similarity of the relative height of the top and bottom plates or the unclear boundaries of the top and bottom plates, a pitch scan command is generated. The pitch scan command refers to controlling the UAV to adjust the body pitch angle or gimbal pitch angle within the safe altitude range to enhance the observation of the top and bottom plate boundaries.

[0068] when When the distance is less than the preset safe distance of the UAV, or when it is necessary to verify the relationship between the observation direction of the current measurement segment and the historical measurement segment, a reverse lookback measurement command is generated. The reverse lookback measurement command means controlling the UAV to look back and collect data in the direction corresponding to the historical measurement segment at the current position, so as to verify whether the local visual feature distribution, laser ray intersection relationship and inertial heading accumulation are consistent.

[0069] The aforementioned lateral scan command, pitch scan command, and reverse lookback measurement command are all verification flight action commands; verification flight action commands are used by the S320 to acquire verification measurement segments.

[0070] S320 controls the UAV to perform short-range measurements according to the verification flight action instructions generated by S310. During the entire process of performing the short-range measurement action, the above-mentioned redundant verification mechanism runs in parallel and monitors the obstacle avoidance distance in real time. Without triggering the emergency braking threshold, verification laser point cloud data, verification visual image data and verification inertial attitude data are collected simultaneously.

[0071] The verification travel distance is jointly limited by the cross-sectional width corresponding to the suspected closed-loop position and the safe flight boundary of the UAV, and is determined according to the following formula:

[0072] in, Indicates the distance moved for verification. This indicates the preset scaling factor. This indicates the cross-sectional width corresponding to the suspected closed-loop location. This indicates the safe flight boundaries for drones as defined by S310.

[0073] The cross-sectional width is derived from S220. Derived from the calculation results of S310, this avoids the UAV entering the collision risk area when performing verification measurements. Verification laser point cloud data, verification visual image data, and verification inertial attitude data are still processed according to the time synchronization rules of S110, the coordinate transformation rules of S120, and the sensor credibility calculation rules of S130, and a "verification" mark is added to the data status label.

[0074] After encapsulation, a verification measurement segment is formed. A verification measurement segment refers to a data segment collected by the UAV near the suspected closed-loop position according to the verification flight action command and encapsulated according to a unified time identifier. It includes at least the verification flight action type, verification start and end time, verification start and end pose, verification laser point cloud data, verification visual image data, verification inertial attitude data, verification cross-sectional contour, verification top and bottom plate relative height, verification side curvature, and verification laser ray set.

[0075] For example, when the suspected closed-loop triggering cause is due to the similar curvature of the left and right sidewalls, and the distances of the UAV to the left sidewall, right sidewall, top plate, and bottom plate are all greater than the preset safety distance, the system performs a lateral scan and encapsulates the data obtained during the lateral scan into a verification measurement segment. The verification measurement segment is used for S330 to match and correspond with historical measurement segments.

[0076] S330. Match the verification measurement segments generated in S320 with the historical measurement segments in the suspected closed-loop records. During matching, use the current robust pose and historical robust pose in the suspected closed-loop records as the initial alignment relationship, transform the verification measurement segments to the underground goaf measurement coordinate system, and establish a correspondence relationship of similar features according to the cross-section position, sampling direction, and verification flight action type.

[0077] For the cross-sectional profile, the correspondence between the verified cross-sectional profile and the historical cross-sectional profile is established based on the spatial position and boundary sequence of the cross-sectional boundary points; for the relative height of the top and bottom plates, the correspondence between the verified relative height of the top and bottom plates and the historical relative height of the top and bottom plates is established based on the height difference at the same cross-sectional position; for the sidewall curvature, the correspondence is established based on the changing trend of the left and right sidewall boundary points along the cross-sectional direction; for the laser ray intersection relationship, the correspondence is established based on the laser emission direction, hit position, and miss direction.

[0078] The difference in relative height between the top and bottom plates is calculated using the following formula:

[0079] in, This represents the difference in the relative height matching between the top and bottom plates between the verified measurement segment and the historical measurement segment. This indicates the relative height of the top and bottom plates in the verification measurement segment. This indicates the relative height of the top and bottom plates in a historical measurement segment.

[0080] The above correspondence is used to generate verification matching results. Verification matching results refer to the matching data set formed by matching the verification measurement segments with historical measurement segments with similar features. It includes at least the matching results of cross-section change sequence, the matching difference of the relative height of the top and bottom plates, the matching results of sidewall curvature change, the matching results of cumulative inertial heading, the matching results of laser ray intersection relationship, the matching results of visual feature review, and the matching credibility.

[0081] The matching reliability is determined based on the matching reliability of the cross-sectional profile, the matching reliability of the relative height of the top and bottom plates, the matching reliability of the sidewall curvature, and the matching reliability of the laser beam intersection relationship. This serves as the reliability basis for generating the closed-loop counter-evidence factor in S410. Therefore, the verification matching results connect the suspected closed-loop record of S230 and the verification measurement segment of S320, providing calculable input for the subsequent construction of the closed-loop counter-evidence chain in S410.

[0082] S4 specifically includes the following sub-steps: S410: Using the verification matching results generated in S330 as input, read the following data: cross-sectional change sequence matching results, top and bottom plate relative height matching difference, sidewall curvature change matching results, inertial heading cumulative amount matching results, laser ray intersection relationship matching results, visual feature review matching results, and matching credibility. Then, convert the above matching data into a closed-loop rebuttal factor. The closed-loop rebuttal factor refers to the judgment quantity formed by the difference in similar features between the verification measurement segment and the historical measurement segment, used to characterize whether there is a spatial contradiction at the suspected closed-loop location.

[0083] Specifically, a cross-section sequence counter-evidence factor is generated based on the cross-section change sequence matching results; a top and bottom plate height counter-evidence factor is generated based on the top and bottom plate relative height matching difference; a side flank curvature counter-evidence factor is generated based on the side flank curvature change matching results; an inertial heading counter-evidence factor is generated based on the inertial heading cumulative amount matching results; a ray intersection counter-evidence factor is generated based on the laser ray intersection relationship matching results; and a visual retrospective counter-evidence factor is generated based on the visual feature retrospective matching results.

[0084] The cumulative difference in inertial heading is calculated using the following formula:

[0085] in, Indicates the cumulative difference in inertial heading. This represents the cumulative inertial heading corresponding to the verification measurement segment. This represents the cumulative inertial heading corresponding to a historical measurement segment.

[0086] The cumulative inertial heading corresponding to the verification measurement segment comes from the verification inertial attitude data collected by S320, while the cumulative inertial heading corresponding to the historical measurement segment comes from the historical measurement segment in the suspected closed-loop record of S230. If the two measurement segments are indeed at the same spatial location, the sequence of cross-section changes, the relative height of the top and bottom plates, the change in sidewall curvature, the cumulative inertial heading, and the intersection relationship of the laser rays should be consistent within the allowable error. If any of the above types of data shows a difference that cannot be explained by sensor error, the corresponding closed-loop disprovenance factor is used by S420 to determine whether the disprovenance is valid.

[0087] S420. Perform a consistency check on the closed-loop counter-evidence factors formed in S410, and determine the counter-evidence status of each closed-loop counter-evidence factor according to the preset tolerance threshold. The preset tolerance threshold is determined based on the UAV measurement accuracy, lidar ranging error, visual camera calibration error, inertial measurement unit attitude error, and allowable measurement error in the goaf area, and remains consistent throughout the same measurement mission.

[0088] For example, when a drone is equipped with a lidar with a ranging accuracy of ±3cm, the preset altitude tolerance can be set to 0.15 meters; the preset curvature tolerance for the sidewall curvature can be set to 0.05 meters. -1 To accommodate the cumulative drift of the inertial navigation system, a preset heading difference threshold of 3 degrees is set. If the data difference between the verified measurement segment and the historical segment exceeds the corresponding value, the corresponding closed-loop disproven factor is valid. =1).

[0089] The state of a single proof by contradiction is determined by the following formula:

[0090] in, Indicates the first The contradictory state of a quasi-closed-loop contradictory factor. The category number represents the closed-loop proof of contradiction factor. Indicates the first Matching difference of quasi-closed-loop proof factors, Indicates the first The preset tolerance threshold for the closed-loop proof factor; when When the value is 1, it indicates that the proof of this type of closed-loop contradiction factor is true. When the value is 0, it indicates that the proof of contradiction for this type of closed-loop proof factor is invalid.

[0091] Specifically, when the sequence of cross-section changes is inconsistent with the historical measurement segments, the cross-section sequence disproving factor is valid; when the difference in the relative height matching of the top and bottom plates is greater than the preset height tolerance, the top and bottom plate height disproving factor is valid; when the sidewall curvature changes in opposite directions and the curvature difference is greater than the preset curvature tolerance, the sidewall curvature disproving factor is valid; when the cumulative difference in inertial heading is greater than the preset heading difference threshold, the inertial heading disproving factor is valid; when the position where the laser hit point should be formed in the verification measurement segment is shown as a non-hit direction in the historical measurement segment, and the difference exceeds the preset ray intersection tolerance, the ray intersection disproving factor is valid.

[0092] The closed-loop counter-evidence factors are organized sequentially according to the cross-sectional change order, the relative height of the top and bottom plates, the sidewall curvature, the cumulative inertial heading, and the laser beam intersection relationship to form a closed-loop counter-evidence chain. A closed-loop counter-evidence chain refers to a set of counter-evidence factors arranged in the above order, each with a counter-evidence status, matching difference, preset tolerance threshold, and matching confidence level. This set is used by S430 to comprehensively determine whether a suspected closed-loop location is a pseudo-similar closed loop.

[0093] S430. Based on the closed-loop counter-evidence chain formed in S420, calculate the comprehensive counter-evidence score of the suspected closed-loop position and generate the closed-loop judgment result.

[0094] The score for the comprehensive proof by contradiction is calculated according to the following formula:

[0095] in, This represents the score for the comprehensive proof by contradiction. This indicates the number of closed-loop proof factors. Indicates the first The weights of the quasi-closed-loop proof factors Indicates the first The contradictory state of a quasi-closed-loop contradictory factor. Indicates the first The matching credibility of the closed-loop counter-evidence factor; the matching credibility comes from the verification matching results generated by S330.

[0096] Inconsistent cross-sectional change sequence and non-corresponding laser beam intersection relationships are defined as strong counter-evidence factors, while differences in relative height between the top and bottom plates, differences in sidewall curvature, differences in cumulative inertial heading, and differences in visual feature replay are defined as weak counter-evidence factors. When a strong counter-evidence factor exists and the counter-evidence is established, or when the comprehensive counter-evidence score is not lower than the preset counter-evidence score threshold, the suspected closed loop position is determined to be a pseudo-similar closed loop. When no strong counter-evidence factor exists and the counter-evidence is established, and the comprehensive counter-evidence score is lower than the preset counter-evidence score threshold, the suspected closed loop position is determined to be a valid closed loop.

[0097] False similarity loop closure refers to two local map units that meet the loop closure candidate conditions in terms of structural similarity, but exhibit spatial inconsistencies after verification by a loop closure counter-proof chain, and therefore cannot be added as valid loop closure constraints to the global map optimization. The loop closure determination result should include at least the suspected loop location identifier, determination type, valid loop closure counter-proof factor, comprehensive counter-proof score, verification matching result number participating in the determination, a marker indicating whether it is allowed to be added to the global map optimization, and the index of verification measurement segments required for subsequent local pose correction.

[0098] Therefore, the closed-loop determination result is directly used by S510 to determine whether a valid closed-loop constraint is generated or not, and by S520 to call the verification measurement segment in the region corresponding to the pseudo-similar closed loop for local pose correction.

[0099] S5 specifically includes the following sub-steps: S510. Based on the loop closure determination result generated in S430, classify the suspected loop closure records formed in S230. Read the suspected loop closure location identifier, determination type, valid loop closure disproving factor, comprehensive disproving score, whether to add global map optimization marker, and verification measurement segment index from the loop closure determination result; when the determination type is valid loop closure, generate valid loop closure constraints based on the current robust pose, historical robust pose, the relative pose relationship between the two, structural similarity, feature credibility, verification matching result number, and loop closure determination result.

[0100] Valid closed-loop constraints refer to closed-loop constraint information that, after verification by the closed-loop counter-evidence chain, has no spatial contradictions and allows inclusion in global map optimization. These constraints limit the relative positional relationship between the current measurement segment and historical measurement segments in the underground goaf measurement coordinate system. When the determination type is pseudo-similar closed loop, the corresponding suspected closed-loop record is not generated as a valid closed-loop constraint, and the corresponding closed-loop constraint addition status is set to 0. Simultaneously, the verified measurement segment, the valid closed-loop counter-evidence factor, the comprehensive counter-evidence score, the verified measurement segment index, and the local pose correction range marker are retained as the basis for S520 to perform local pose correction.

[0101] The closed-loop constraint state is determined according to the following formula:

[0102] in, This indicates the state where closed-loop constraints are applied. This indicates the closed-loop determination result output by S430; when When the value is 1, the corresponding effective closed-loop constraint is added to the global map optimization. When the value is 0, no corresponding closed-loop constraint is generated, and the corresponding verification measurement segment is transferred to the local pose correction process.

[0103] Therefore, S510 does not delete suspected closed-loop records, but uses valid closed loops for global constraints, and retains the verification measurement segments in pseudo-similar closed loops as the basis for local correction, thus avoiding erroneous closed loops from polluting the global map.

[0104] S520: Using the robust pose trajectory generated by S210 as the object to be corrected, and taking the effective closed-loop constraints generated by S510, the local map units generated by S210, the verification measurement segments generated by S320, and the verification matching results generated by S330 as the optimization basis, global map optimization and local pose correction are performed.

[0105] During the global map optimization process, the continuous motion relationship between adjacent robust poses is used as the motion continuity constraint, the effective closed loop constraint is used as the global closure constraint, the laser point cloud registration relationship between local map units is used as the spatial consistency constraint, and the cross-sectional contour, the relative height of the top and bottom plates, the side curvature, and the intersection relationship of laser rays in the verification measurement segment are used as the cross-sectional constraint, so that the corrected pose trajectory simultaneously satisfies continuous motion, effective closed loop, and local cross-sectional consistency.

[0106] The global map optimization error is expressed by the following formula:

[0107] in, This indicates the global map optimization error. This represents the continuous motion error between adjacent robust poses. This indicates the effective closed-loop constraint error. This indicates the cross-sectional constraint error corresponding to the verification measurement segment; This represents the spatial consistency constraint error between local map units; by reducing This yields a high-precision positioning trajectory.

[0108] For regions identified as pseudo-similar loop closures by S430, a preset number of local map units are extended forward and backward along the robust pose trajectory, centered on the corresponding current measurement segment, to form a local pose correction range. The local pose correction range refers to the interval of the trajectory to be corrected formed on the robust pose trajectory, centered on the current measurement segment corresponding to the pseudo-similar loop closure.

[0109] Within this range, the local pose correction is calculated based on the cross-sectional profile, relative height of the top and bottom plates, sidewall curvature, and laser beam intersection relationship in the verification measurement segment, and the original robust pose is locally corrected:

[0110] in, This indicates the pose after local correction. This represents the robust pose generated by S210. This represents the local pose correction amount calculated based on the verified measurement segment.

[0111] High-precision positioning trajectory refers to the continuous pose trajectory of a UAV obtained by re-optimizing it using effective closed-loop constraints, verification measurement segments, and local map units after eliminating pseudo-similarity closed-loop constraints. Subsequently, the high-precision positioning trajectory is reprojected and registered with laser point cloud data and visual image data to obtain a corrected 3D map of the mined-out area.

[0112] S530. Based on the high-precision positioning trajectory obtained in S520 and the corrected 3D map of the goaf, generate the fused SLAM measurement results for the goaf. The corrected 3D map of the goaf refers to the spatial map of the underground goaf formed by reprojecting and registering the high-precision positioning trajectory with laser point cloud data and visual image data.

[0113] First, the outer boundary surface formed by continuous effective point clouds is extracted from the corrected 3D map of the goaf to obtain the goaf boundary. Second, cross-sectional point clouds approximately perpendicular to the UAV's travel direction are extracted at preset intervals along a high-precision positioning trajectory. Cross-sectional contours are generated based on the effective boundary points of the roof, floor, left side, and right side in the cross-sectional point clouds. Third, the cavity volume is calculated based on the cross-sectional area and spacing corresponding to adjacent cross-sectional contours.

[0114] in, Indicates the volume of the cavity. Indicates the first The cross-sectional area corresponding to each cross-sectional profile. Indicates the first The cross-sectional area corresponding to each cross-sectional profile. Indicates the first The profile of the first cross section and the first The spacing between the cross-sectional profiles along a high-precision positioning trajectory. Indicates the number of cross-sectional profiles. A positive integer representing the sequence number of the cross-sectional profile.

[0115] The coordinates of the dangerous area are determined based on the abnormal areas of the roof boundary, the collapse accumulation boundary, the abrupt change area of ​​the cross section, and the abnormal boundary areas confirmed by multi-angle verification measurement, and are uniformly linked to the underground goaf measurement coordinate system; the abnormal areas of the roof boundary are determined by the abrupt change in the height of the effective boundary point of the roof in the adjacent cross section, the discontinuity of the roof point cloud, or the local subsidence of the roof boundary surface.

[0116] Finally, the boundaries, cross-sectional contours, cavity volumes, coordinates of hazardous areas, closed-loop determination results, effective closed-loop constraints, suspected closed-loop records identified as pseudo-similar closed loops, verification measurement segment indexes, and local pose correction records of the goaf are associated and stored to form the goaf fusion SLAM measurement results. The goaf fusion SLAM measurement results are used to characterize the spatial measurement results of the underground goaf after the influence of pseudo-similar closed loops has been eliminated.

[0117] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0118] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0119] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0120] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A UAV-based fusion SLAM measurement method for underground goaf areas, characterized in that, Includes the following steps: S1. Control the UAV to enter the underground mining area and simultaneously collect laser point cloud data, visual image data and inertial attitude data through lidar, vision camera and inertial measurement unit. After time synchronization, coordinate transformation and validity screening, a fused measurement data sequence is formed. S2. Based on the fused measurement data sequence, robust pose estimation is performed to generate robust pose trajectory, and local map units are constructed. The cross-sectional contour, relative height of top and bottom plates, side curvature and local visual features are extracted, and suspected closed-loop records are generated by machine learning similarity discrimination model. S3. Determine the safe flight boundary of the UAV based on the suspected closed-loop record, establish a redundancy verification mechanism based on real-time perception data, generate verification flight action commands, control the UAV to perform lateral scanning, pitch scanning or reverse review, form verification measurement segments, and match them with historical measurement segments to generate verification matching results. S4. Generate closed-loop counter-evidence factors based on the verification matching results, construct a closed-loop counter-evidence chain, and output the closed-loop determination result of valid closed loop or pseudo-similar closed loop based on the closed-loop counter-evidence chain.

2. The UAV-based fusion SLAM measurement method for underground mining subsidence areas according to claim 1, characterized in that, Also includes: S5. Generate effective closed-loop constraints or not generate corresponding closed-loop constraints based on the closed-loop determination results. Use the effective closed-loop constraints, local map units and verification measurement segments to correct the robust pose trajectory, obtain the high-precision positioning trajectory and the corrected three-dimensional map of the goaf, and generate the fusion SLAM measurement results of the goaf.

3. The UAV-based fusion SLAM measurement method for underground mining subsidence areas according to claim 1, characterized in that, S1 specifically includes: The system reads the preset measurement path, entry reference coordinates, initial attitude of the UAV, and installation parameters of each sensor. Using the sampling time of the inertial measurement unit as the main time axis, it synchronizes the laser point cloud data, visual image data, and inertial attitude data to form initial synchronized acquisition data. A measurement coordinate system for the underground goaf area is established based on the entrance reference coordinates and the initial attitude of the UAV. The laser point cloud data is converted to this coordinate system, and invalid frames, frames with missing timestamps, and abnormal data from the sensors are removed to obtain effective synchronous measurement data.

4. The UAV-based fusion SLAM measurement method for underground mining subsidence areas according to claim 3, characterized in that, It also includes: encapsulating effective synchronous measurement data according to a unified time identifier, recording the UAV sampling location, point cloud density, image clarity, inertial attitude change, data status markers and sensor reliability, forming a fused measurement data sequence.

5. The UAV-based fusion SLAM measurement method for underground mining subsidence areas according to claim 1, characterized in that, S2 specifically includes: Using the fused measurement data sequence as input, laser point cloud data, visual image data, inertial attitude data, UAV sampling position and sensor reliability are read to perform inertial prediction, point cloud registration correction and visual feature tracking correction, generate robust pose trajectory and construct local map units; The cross-sectional contour, relative height of top and bottom plates, cross-sectional width, sidewall curvature, tunnel width-to-height ratio, local visual features, and laser point cloud distribution features are extracted from local map units, input into a machine learning similarity discrimination model, and combined with weight formulas to obtain structural similarity and feature credibility. Non-adjacent local map units are used as candidate comparison objects. When the structural similarity, feature credibility, and path interval all meet the threshold conditions, a suspected closed loop record is generated.

6. The UAV-based fusion SLAM measurement method for underground mining subsidence areas according to claim 1, characterized in that, S3 specifically includes: After generating a suspected closed-loop record, the current measurement segment, historical measurement segment, current robust pose, historical robust pose and suspected closed-loop triggering cause are read. Combined with the cross-section width, relative height of top and bottom plates, side curvature, local visual feature matching distribution and laser point cloud set, the safe flight boundary of the UAV is determined. Combined with the real-time ranging data of the UAV, a redundant verification mechanism for collision avoidance is constructed, and a verification flight action command is generated. The drone is controlled to perform short-range measurements according to the verification flight command. During the execution, the obstacle avoidance distance is monitored in real time through a redundancy verification mechanism. Simultaneously, verification laser point cloud data, verification visual image data, and verification inertial attitude data are collected and packaged into verification measurement segments according to a unified time identifier.

7. The UAV-based fusion SLAM measurement method for underground goaf areas according to claim 6, characterized in that, Also includes: The verification measurement segments and historical measurement segments are matched by establishing a correspondence of similar features based on cross-sectional location, sampling direction, and action type, and verification matching results are generated.

8. The UAV-based fusion SLAM measurement method for underground mining subsidence areas according to claim 1, characterized in that, S4 specifically includes: Using the verification matching results as input, the sequence of cross-section changes, the difference in relative height between the top and bottom plates, the change in sidewall curvature, the cumulative amount of inertial heading, the intersection relationship of laser rays, the visual feature review of matching data and the matching credibility are read to generate a closed-loop counter-evidence factor. The consistency of each closed-loop counter-evidence factor is checked according to the preset tolerance threshold to determine the counter-evidence status, and a closed-loop counter-evidence chain is formed according to the sequence of cross-sectional changes, the relative height of the top and bottom plates, the curvature of the side walls, the cumulative amount of inertial heading, and the intersection relationship of laser rays. The comprehensive counter-proof score is calculated based on the closed-loop counter-proof chain. If a strong counter-proof factor is found or the comprehensive counter-proof score reaches the threshold, it is determined to be a pseudo-similar closed loop; otherwise, it is determined to be a valid closed loop, and a closed loop determination result is generated.

9. The UAV-based fusion SLAM measurement method for underground mining subsidence areas according to claim 2, characterized in that, S5 specifically includes: Based on the closed-loop determination results, suspected closed-loop records are processed. When a valid closed loop is found, a valid closed-loop constraint is generated and added to the global map optimization. When a pseudo-similar closed loop is found, no closed-loop constraint is generated. The verification measurement segments and local pose correction range markers are retained. Using robust pose trajectory as the object to be corrected, global map optimization and local pose correction are performed by utilizing effective closed-loop constraints, local map units, verification measurement segments, and verification matching results to obtain high-precision positioning trajectory and corrected 3D map of goaf area.

10. The UAV-based fusion SLAM measurement method for underground goaf areas according to claim 9, characterized in that, It also includes: extracting the goaf boundary, cross-sectional contour, cavity volume and dangerous area coordinates based on high-precision positioning trajectory and corrected 3D map of goaf, and associating the closed loop judgment results, effective closed loop constraints, suspected closed loop records judged as pseudo-similar closed loops, verification measurement segment index and local pose correction records to form the goaf fusion SLAM measurement results.

Images